Intelligent classification of lung cancer pathology images through comparative morphological feature learning

Area of Science:

• Computational pathology
• Medical image analysis
• Artificial intelligence in oncology

Background:

• Accurate lung cancer pathology image classification is crucial for diagnosis and treatment.
• Challenges include complex cellular structures and limited labeled data, hindering model development.

Purpose of the Study:

• To improve lung cancer pathology image classification by incorporating unlabeled data.
• To leverage comparative learning techniques for enhanced model training.

Main Methods:

• Utilized a ResNet50 encoder with deformable and dynamic convolutions for feature extraction.
• Integrated confidently classified unlabeled images with labeled data for training.
• Employed farthest and nearest neighbor contrastive learning for a challenging learning environment.

Main Results:

• Achieved significant improvements in lung cancer image classification accuracy.
• Demonstrated robustness of the method, especially with limited labeled data.

Conclusions:

• Comparative learning with labeled and unlabeled data, plus advanced convolutions, enhances lung cancer image classification.
• Presents a practical solution for accurate and efficient oncological diagnostic tools.